Radio frequency interference identification method and system based on astronomical time-frequency images

Through the MS-TransUNet network model based on multi-scale feature fusion, combined with Transformer and U-Net architecture, a multi-scale residual position perception attention module is introduced, which solves the problem of limited RF interference identification efficiency and accuracy in the prior art, and realizes the implementation of the implementation of the problem of the problem of limited RF interference recognition efficiency and accuracy. Efficient and accurate identification in astronomical time-frequency images.

CN118736310BActive Publication Date: 2024-11-22KUNMING UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410871662.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-01
Publication Date
2024-11-22
Estimated Expiration
2044-07-01

AI Technical Summary

Technical Problem

The existing RF interference identification methods are difficult to fully and accurately identify multiple RF interference signals when processing astronomical time-frequency images, especially in the astronomical environments with limited data sets and complex data sets, resulting in limited identification efficiency and accuracy.

Method used

Using the MS-TransUNet network model based on multi-scale feature fusion composed of encoder, Branch branch and decoder, combined with Transformer and U-Net architecture, a multi-scale residual position perception attention module is introduced to extract high-dimensional features and The encoder extracts local features, integrates spatial information and higher-order semantic information, and outputs radio frequency interference mask images.

Benefits of technology

Achieves fast network convergence speed and reasonable performance, can fully and accurately identify RF interference, with lightweight and high efficiency, and is suitable for the situation of limited data sets, improving the comprehensiveness and accuracy of RF interference recognition .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118736310B_ABST
    Figure CN118736310B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for identifying radio frequency interference based on astronomical time-frequency images, including: obtaining astronomical time-frequency images; using the astronomical time-frequency images as inputs of encoders and Branch branches to obtain encoder feature maps and high-dimensional feature maps respectively; using the encoder feature maps as inputs of decoders, introducing a multi-scale residual position perception attention module in the decoder to extract feature map information, splicing the output of the fourth upsample in the decoder with the output of the Branch branch to obtain a spliced ​​feature map, and integrating spatial information with high-order semantic information; the spliced ​​feature map is subjected to one convolution operation and then predicted by a segmentation head to classify each pixel. Compared with existing similar methods, it has a faster network convergence speed, reasonable network performance, and can comprehensively and accurately identify RFI.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a radio frequency interference identification method and system based on astronomical time-frequency images, and belongs to the field of image processing based on computer vision. Background Art

[0002] Astronomy is a science that requires high precision and clarity, and the presence of radio frequency interference may have a negative impact on astronomical images and data. Therefore, radio frequency interference identification has become an important field in astronomical research. Radio Frequency Interference (RFI) refers to electromagnetic radiation from artificial radio frequency sources, such as signals generated by radio communications, radars, satellite communications, etc., which may interfere with astronomical observation equipment to capture and analyze weak signals from the universe. This interference may cause noise, distortion or data offset in astronomical images, seriously affecting astronomers' research and understanding of the universe. Radio frequency interference can have many sources: communication facilities: including ground radio stations, satellite communications, mobile communication networks, etc.; aircraft and satellites: aircraft, satellites and other communication equipment may generate radio frequency interference; artificial sources: human-made radio frequency sources such as artificial satellites and space stations may also become interference sources. Radio frequency interference may have many effects on astronomical images and data, such as signal distortion: interference signals are mixed with celestial body signals, resulting in unclear signal characteristics; increased noise: interference signals increase the background noise level and reduce the signal-to-noise ratio of the data; data offset: interference signals may cause astronomical data to offset, affecting the accuracy of measurement and analysis.

[0003] As a key task in astronomical research, the development and application of radio frequency interference identification is crucial to ensuring the quality and accuracy of astronomical data. Currently, common radio frequency interference identification methods include spectrum analysis, spatial filtering, signal processing algorithms, time-frequency analysis, and machine learning technology. In recent years, the application of machine learning technology in radio frequency interference identification has gradually increased. By training machine learning models, different types of radio frequency interference signals can be identified, thereby realizing automated interference identification. Therefore, by continuously improving identification methods, astronomers can better cope with the challenges of radio frequency interference and thus explore the mysteries of the universe more deeply.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] The present invention provides a method and system for identifying radio frequency interference based on astronomical time-frequency images, constructs an MS-TransUNet network model based on multi-scale feature fusion composed of an encoder, a Branch branch, and a decoder, and realizes the radio frequency interference identification of astronomical time-frequency images corresponding to real observation data.

[0006] The technical solution of the present invention is:

[0007] According to a first aspect of the present invention, a method for identifying radio frequency interference based on astronomical time-frequency images is provided, comprising: acquiring astronomical time-frequency images; constructing an MS-TransUNet network model based on multi-scale feature fusion composed of an encoder, a Branch branch, and a decoder; using the astronomical time-frequency image as the input of the encoder and the Branch branch, and obtaining an encoder feature map and a high-dimensional feature map respectively; using the encoder feature map as the input of the decoder, introducing a multi-scale residual position-aware attention module in the decoder to extract feature map information, splicing the output of the fourth upsample in the decoder with the output of the Branch branch to obtain a spliced ​​feature map, and integrating spatial information with high-order semantic information; the spliced ​​feature map is subjected to one convolution operation and then predicted by a segmentation head for each pixel category to output a radio frequency interference mask image; wherein the categories include: interference and non-interference.

[0008] The Branch specifically includes: performing an addition operation on the outputs of the input astronomical time-frequency image after being respectively operated by a 3×3 convolution kernel and a 1×1 convolution kernel, and then performing batch normalization and ReLU operations to obtain a high-dimensional feature map.

[0009] The encoder and decoder of the MS-TransUNet network model based on multi-scale feature fusion are based on the encoder and decoder of the TransUNet network as the framework. In the encoder, the Transformer block adopts a Transformer layer; and in the decoder, a multi-scale residual position-aware attention module is introduced.

[0010] The encoder comprises a convolution block and a Transformer block; wherein the convolution block extracts local features of the input astronomical time-frequency image to obtain local feature maps of radio frequency interference targets of three different scales; the Transformer block encodes the local feature map of the radio frequency interference target of the third scale and extracts features from it to obtain an encoder feature map.

[0011] The decoder introduces a multi-scale residual position-aware attention module after the first upsampling.

[0012] The multi-scale residual position-aware attention module specifically comprises: extracting feature information of the multi-scale residual position-aware attention module input image at different scales by setting "expanded convolution with different void rates + average pooling"; concatenating the results of the expanded convolution and the average pooling, and then performing dimensionality reduction through a 1×1 convolution; performing an addition operation on the original input image of the multi-scale residual position-aware attention module and the feature map after dimensionality reduction, and then performing a channel shuffling operation on the feature map after batch normalization and ReLU operation; performing global average pooling on the feature map after the channel shuffling operation in the horizontal and vertical directions respectively, and obtaining two The feature maps in the horizontal and vertical directions are spliced ​​and then processed by 1×1 convolution, batch normalization and nonlinear activation. The processed feature maps are divided into two parts in the channel dimension: they are respectively processed by 1×1 convolution conv and then by Sigmoid activation function to obtain weights; the "output obtained by multiplying the feature map after the channel shuffle operation with the two weights" and the "output of adding the original input image of the multi-scale residual position-aware attention module multiplied by the preset coefficient and the feature map after the channel shuffle operation multiplied by the preset coefficient" are added as the output of the multi-scale residual position-aware attention module.

[0013] The MS-TransUNet network model based on multi-scale feature fusion also includes training and testing steps, specifically: converting real observation data into astronomical time-frequency images, and using a labeling + correction method to produce real mask label images to construct a data set; randomly dividing the data set into a training set, a validation set and a test set according to a preset ratio; using the Transformer enhancement technology to expand the training set to obtain an expanded training set; training and verifying the MS-TransUNet network model based on multi-scale feature fusion based on the expanded training set and the validation set to obtain a trained MS-TransUNet network model based on multi-scale feature fusion; using the astronomical time-frequency images in the test set as inputs of the trained MS-TransUNet network model based on multi-scale feature fusion to obtain a radio frequency interference mask image.

[0014] According to a second aspect of the present invention, a radio frequency interference recognition system based on astronomical time-frequency images is provided, comprising: an acquisition module for acquiring astronomical time-frequency images; an identification module for constructing an MS-TransUNet network model based on multi-scale feature fusion composed of an encoder, a Branch branch, and a decoder; taking the astronomical time-frequency images as inputs of the encoder and the Branch branch, and obtaining encoder feature maps and high-dimensional feature maps respectively; taking the encoder feature maps as inputs of the decoder, introducing a multi-scale residual position-aware attention module in the decoder to extract feature map information, splicing the output of the fourth upsample in the decoder with the output of the Branch branch to obtain a spliced ​​feature map, and integrating spatial information with high-order semantic information; the spliced ​​feature map is subjected to one convolution operation and then predicted by a segmentation head to classify each pixel, so as to output a radio frequency interference mask image; wherein the categories include: interference and non-interference.

[0015] According to a third aspect of the present invention, a processor is provided, the processor being configured to execute an operation, the operation comprising executing any one of the above-mentioned methods for identifying radio frequency interference based on astronomical time-frequency images.

[0016] According to a fourth aspect of the present invention, a computer-readable storage medium is provided for storing a computer program; wherein, when the computer program is executed by a processor, the steps of any one of the above-mentioned methods for identifying radio frequency interference based on astronomical time-frequency images are implemented.

[0017] The beneficial effects of the present invention are as follows: the present invention constructs an MS-TransUNet network model based on multi-scale feature fusion, which is composed of an encoder, a Branch branch, and a decoder. The network combines the ideas of the transformer and u-net architectures, uses the branch branch as a high-dimensional feature extraction part for extracting interference in the time-frequency image, and the multi-scale residual position perception attention module extracts low-dimensional feature information in the feature image. Compared with the existing similar methods, it has a faster network convergence speed, reasonable network performance, and can comprehensively and accurately identify RFI. At the same time, the present invention considers the lightweight of the network, sets fewer parameters and a reasonable number of training times. In the case of a limited data set, the present invention expands the data set by online rotation and cropping of the data during data preprocessing to ensure the efficiency of learning. This also means that the network of the present invention has the characteristics of both lightweight and high efficiency. Compared with the existing deep learning RFI recognition method, the recognition possibility of more types of RFI is increased, and the recognition of RFI is more comprehensive. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flowchart of the present invention;

[0019] Figure 2This is a structural diagram of the MS-TransUNet network model based on multi-scale feature fusion proposed in the present invention;

[0020] Figure 3 This is the Branch structure diagram of the present invention;

[0021] Figure 4 This is a structural diagram of the multi-scale residual position-aware attention module of the present invention;

[0022] Figure 5 This is a pulsar astronomical time-frequency image converted based on real observation data;

[0023] Figure 6 Based on Figure 5 The produced true mask label image;

[0024] Figure 7 The MS-TransUNet of the present invention is Figure 5 The recognition result diagram of

[0025] Figure 8 For U-net Figure 5 The recognition result graph of

[0026] Fig. 9 For R-Net Figure 5 The recognition result graph of

[0027] Fig.10 For RFI-Net Figure 5 The recognition result graph of

[0028] Fig.11 For DSC Figure 5 The recognition result diagram. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments in this application and the features in the embodiments can be combined with each other arbitrarily without conflict.

[0030] Example 1: Figure 1-11As shown, according to the first aspect of an embodiment of the present invention, a method for identifying radio frequency interference based on astronomical time-frequency images is provided, including: obtaining astronomical time-frequency images; constructing a MS-TransUNet network model based on multi-scale feature fusion composed of an encoder, a Branch branch, and a decoder; using the astronomical time-frequency image as the input of the encoder and the Branch branch, respectively obtaining an encoder feature map and a high-dimensional feature map; using the encoder feature map as the input of the decoder, introducing a multi-scale residual position-aware attention module in the decoder to extract feature map information, splicing the output of the fourth upsample in the decoder with the output of the Branch branch to obtain a spliced ​​feature map, and integrating spatial information with high-order semantic information; the spliced ​​feature map is subjected to one convolution operation and then predicted by a segmentation head for each pixel category to output a radio frequency interference mask image; wherein the categories include: interference and non-interference. According to the specific burst characteristics of radio frequency interference, a Branch branch is designed between the encoder and the decoder to save the texture and shape information in the image, so that the network can combine more low-dimensional image feature information after upsampling Upsample to achieve a more accurate interference recognition effect.

[0031] Furthermore, the Branch branch specifically includes: performing an addition operation on the output of the input astronomical time-frequency image after being operated by a 3×3 convolution kernel and a 1×1 convolution kernel respectively, and then performing batch normalization and Relu operations to obtain a high-dimensional feature map. By applying the above technical solution, it can be seen that the 3×3 convolution kernel introduced by the Branch branch can capture the local features of radio frequency interference in the image, such as edges, textures, etc.; by sliding the 3×3 convolution kernel on the image, the network can effectively learn the local structure in the image, which is very important for distinguishing the boundaries and details of different information in the radio frequency interference identification task. On this basis, the subsequent 1×1 convolution can reduce the dimension of the feature map while retaining sufficient boundary and detail information to reduce the amount of calculation and memory consumption.

[0032] Furthermore, the encoder and decoder of the MS-TransUNet network model based on multi-scale feature fusion are based on the encoder and decoder of the TransUNet network as the framework. In the encoder, the Transformer block adopts a Transformer layer; and a multi-scale residual position-aware attention module is introduced in the decoder.

[0033] Furthermore, the encoder includes a convolution block and a Transformer block; wherein the convolution block CNN extracts local features of the input astronomical time-frequency image to obtain local feature maps of the radio frequency interference target at three different scales, which helps to capture the spatial information of the RFI in the image; the Transformer block encodes the local feature map of the radio frequency interference target at the third scale and extracts features from it to obtain an encoder feature map. It should be noted that the Transformer block uses a self-attention mechanism, which can apply self-attention to the local neighborhood of each query pixel in the input third-scale RFI target local feature map, performs tokenization by inputting a flat patch sequence, and uses a trainable linear projection to map the vectorized patches to the potential embedding space, which enables the model to pay attention to the contextual information of the RFI target information.

[0034] Furthermore, the decoder introduces a multi-scale residual position-aware attention module after the first upsampling.

[0035] Furthermore, the decoder includes four upsampling layers (the upsampling layer uses deconvolution to restore the low-resolution feature map to a high-resolution feature map), six convolution operations (five for feature extraction and one for the segmentation head), and in order to accurately locate and identify the shape and position of RFI in the original astronomical signal, a multi-scale residual position-aware attention module is introduced after the first upsampling of the decoder, which enables the network model to focus on the information of interest and emphasizes the shape and position of RFI in the original astronomical signal, thereby improving the segmentation accuracy of RFI in the image. The four upsampling operations used in the model combine the three different scales of the RFI target local feature maps from the encoder, realize the combination of feature maps of different resolutions, and enable the model to recognize RFI features of different scales. The output of the four upsampling and four convolution operations is spliced ​​with the high-dimensional image output by the Branch branch to obtain a spliced ​​feature map, realizing the integration of spatial information and high-order semantic information; the spliced ​​feature map is then output by the segmentation head after one convolution operation and one convolution operation.

[0036] Furthermore, the multi-scale residual position perception attention module specifically comprises: extracting feature information of the multi-scale residual position perception attention module input image at different scales by setting "expanded convolution with different voiding rates + average pooling"; concatenating the results of the expanded convolution and the average pooling, and then performing dimensionality reduction through a 1×1 convolution; performing an addition operation on the original input image of the multi-scale residual position perception attention module and the feature map after dimensionality reduction, and then performing a channel shuffling operation on the feature map after batch normalization and Relu operation. Shuffle to increase the interaction between channels; perform global average pooling on the feature map after the channel shuffle operation in the horizontal and vertical directions to obtain two feature maps, concatenate the feature maps in the horizontal and vertical directions, and then process them through 1×1 convolution, batch normalization and nonlinear activation (such as h_swish), and divide the processed feature map into two parts in the channel dimension: respectively through 1×1 convolution conv, and then through the Sigmoid activation function to obtain the weight; add the "output obtained by multiplying (dot product) the feature map after the channel shuffle operation and the two weights" and the "output of adding the original input image of the multi-scale residual position-aware attention module multiplied by the preset coefficient and the feature map after the channel shuffle operation multiplied by the preset coefficient" as the output of the multi-scale residual position-aware attention module. Among them, the preset coefficient is 0.1. The multi-scale residual position perception attention module specially designed for the identification task of radio frequency interference in the present invention can not only capture feature information of different scales, but also enhance the interaction and feature expression ability between channels, realize the attention to the coordinate information of RFI features at different positions in the image, improve the model's perception and understanding of image semantic information, and enable the model to more accurately locate and identify RFI. Specifically: by setting different hole rate expansion convolutions to extract feature information at different scales, it can be ensured that the number of model parameters will not increase significantly while maintaining the multi-scale receptive field, and the use of channel shuffling operations can alleviate the problem of information exchange between isolated channels caused by expansion convolution; further, the RFI feature position information after the expansion convolution and channel shuffling operations is embedded in the channel attention, so that the model can obtain larger regional information without introducing a large amount of computational overhead; furthermore, the channel attention is decomposed into two parallel one-dimensional feature codes, which can avoid the loss of position information caused by 2D global average pooling, and effectively integrate the spatial position information of RFI; finally, the information integration of each part introduces residual connection to prevent the gradient in the network training process and integrate more RFI feature information. This module uses the idea of ​​multi-scale and considers the relationship between channel and position information at the same time.

[0037] Furthermore, the MS-TransUNet network model based on multi-scale feature fusion also includes training and testing steps, specifically: converting real observation data into astronomical time-frequency images, and using the annotation + correction method to produce real mask label images to construct a data set; randomly dividing the data set into a training set, a validation set and a test set according to a preset ratio; using the Transformer enhancement technology to expand the training set to obtain an expanded training set; specifically, using the Transformer technology to perform random cropping and rotation operations to achieve expansion; training and verifying the MS-TransUNet network model based on multi-scale feature fusion based on the expanded training set and the validation set to obtain a trained MS-TransUNet network model based on multi-scale feature fusion; using the astronomical time-frequency images in the test set as the input of the trained MS-TransUNet network model based on multi-scale feature fusion to obtain a radio frequency interference mask image; further, the astronomical time-frequency images converted from other original astronomical data can be input, the MS-TransUNet network model based on multi-scale feature fusion can be called, and the predicted radio frequency interference mask image can be output. In addition, the predicted RFI mask image can be compared with the real mask label image.

[0038] Furthermore, precision, recall and F2 score are used to evaluate the model in both training and testing processes.

[0039] According to a second aspect of an embodiment of the present invention, a system for identifying radio frequency interference based on astronomical time-frequency images is provided, comprising: an acquisition module for acquiring astronomical time-frequency images; an identification module for constructing an MS-TransUNet network model based on multi-scale feature fusion composed of an encoder, a Branch branch, and a decoder; taking the astronomical time-frequency images as inputs of the encoder and the Branch branch, and obtaining encoder feature maps and high-dimensional feature maps respectively; taking the encoder feature maps as inputs of the decoder, introducing a multi-scale residual position-aware attention module in the decoder to extract feature map information, splicing the output of the fourth upsample in the decoder with the output of the Branch branch to obtain a spliced ​​feature map, and integrating spatial information with high-order semantic information; the spliced ​​feature map is subjected to one convolution operation and then predicted by a segmentation head to classify each pixel, so as to output a radio frequency interference mask image; wherein the categories include: interference and non-interference.

[0040] According to a third aspect of an embodiment of the present invention, a processor is provided, wherein the processor is used to execute an operation, wherein the operation includes executing any one of the above-mentioned methods for identifying radio frequency interference based on astronomical time-frequency images.

[0041] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided for storing a computer program; wherein, when the computer program is executed by a processor, the steps of any one of the above-mentioned methods for identifying radio frequency interference based on astronomical time-frequency images are implemented.

[0042] It should be noted that the source of RF interference may come from many different places, including communication equipment, satellite communications, aircraft, radar, etc. These interference sources may have different frequencies, powers and operating modes, making it more complicated and difficult to identify and distinguish them. In astronomical observations, interference signals are often mixed with celestial signals, resulting in distortion or blurring of celestial signals. This mixing makes it more difficult to distinguish and identify interference signals, and more sophisticated processing methods are required. Astronomical observations are usually carried out in complex and changeable astronomical environments, such as weather changes, earth rotation, atmospheric interference, etc., which may affect the propagation and reception of RF signals. These factors increase the difficulty of identifying and eliminating RF interference. In combination with experimental data, the optional specific implementation process of the present invention is described as follows:

[0043] In this experiment, the present invention sets the batch size to 8, with a total of 300 epochs, and the initial weights of each layer are random. The dice loss is used as the loss function, Adam is used as the optimizer, and the learning rate is set to be automatically adjusted randomly during the training process.

[0044] Step 1. The data set of the present invention is produced by the laboratory. The model is trained and evaluated using real pulsar observation data from the 40-meter radio telescope of the Yunnan Observatory of the Chinese Academy of Sciences from September 2016 to October 2022. The central observation frequency of the experimental data is 2256MHz, the number of phase points is 512, and the data with more than 90 sub-integration channels and more than 128 frequency channels are selected. The sub-integration time includes 10s and 30s, and the frequency channel bandwidth is 1MHz.

[0045] Step 2: First, process the original real pulsar observation data and convert it into astronomical time-frequency images. Then, use AOFlagger 3 to annotate and perform secondary manual correction to produce mask label images.

[0046] Step 3. The experimental data set of this study is made of 1383 observation files of three different pulsars. To ensure the fairness and rationality of the data set division, the present invention randomly divides the training set, validation set and test set in a ratio of 7:1.5:1.5, that is, 70% as training samples, 15% as validation samples, and the remaining 15% as test samples. Among them, 968 files are used for model training, 207 files are used for model verification, and the remaining 208 files are used for model performance evaluation.

[0047] Furthermore, the present invention uses the enhancement method in transformer technology to perform random rotation and cropping operations to expand the diversity of data, improve the learning efficiency of the network, and expand the training set to 290,400 images online.

[0048] It should be noted that any given time-frequency image in the training, validation and test sets contains information about useful pulsar signals and useless RFI signals, and the mask label image corresponding to this image only contains RFI signals;

[0049] During the learning and training process, the MS-TransUNet network model based on multi-scale feature fusion learns the interference information in the original time-frequency image, and uses the mask label as a reference to adjust the learning method and strategy; the verification and testing process uses precision, recall and F2 score to evaluate the model. The specific instructions are as follows:

[0050] The precision rate represents the ratio of the pixels in the time-frequency image that are correctly predicted by the model to the total number of pixels predicted to be RFI. The formula for the precision rate is:

[0051]

[0052] Among them, TP (True Positives) is the number of pixels that the model correctly predicts as RFI, and FP (False Positives) is the number of pixels that the model incorrectly predicts as RFI.

[0053] Recall rate refers to the proportion of samples that are truly positive examples that the model successfully predicts as positive examples. In the RFI identification task, recall rate represents the ratio of pixels in the time-frequency image that the model correctly predicts as RFI to all pixels that are actually predicted as RFI. The formula for recall rate is:

[0054]

[0055] Where TP is the number of pixels that the model correctly predicts as RFI, and FN (False Negative) is the number of pixels that the model fails to correctly predict as RFI.

[0056] The fβ score is the weighted harmonic mean of precision and recall. By adjusting the β value, the relationship between precision and recall can be balanced. The formula for the Fβ score is:

[0057]

[0058] Wherein, β is a weight parameter used to balance precision and recall. In the F2 score, the larger the value of β, the greater the impact on the recall. In the embodiment of the present invention, β is set to 2.

[0059] The results are recorded each time the model is fine-tuned, and the comprehensive indicator F2 score is used as the evaluation indicator, and the model with the best results is saved.

[0060] During the training process, the value of the loss function changes with the progress of the experiment. The values ​​of various evaluation indicators are also changing. Through a random experiment, it can be seen that the loss function value of the MS-TransUNet network is close to 0.4 at the beginning, and gradually decreases as the training process progresses. It tends to be stable at the 230th epoch and finally approaches 0. The precision rate is relatively high but unstable at the beginning, and tends to be stable as the number of training increases; the recall rate is relatively low at the beginning, and it increases rapidly at the beginning of training. It reaches a high level at the 230th epoch as the training process progresses and shows a stable level in subsequent training.

[0061] In order to explore and verify the superiority of the network model of the present invention, the present invention reproduces the currently superior U-Net, RFI-Net, DSC and R-Net, etc. The same experimental platform and data set are used to train and test each model, and the average index of the results of multiple experiments is selected for display (the results of 20 experiments are selected), as shown in Table 1:

[0062] Table 1 Comparative experiment

[0063] Precision Recall F2 score U-Net 0.8929 0.7251 0.7534 RFI-Net 0.8792 0.7457 0.7692 DSC 0.8479 0.6922 0.7186 R-Net 0.8319 0.5574 0.5936 MS-TransUNet 0.8129 0.8016 0.8039

[0064] From the data shown in Table 1, it can be seen that precision and recall have mutually inhibiting performance, and one rises and the other must fall. Precision represents the precision rate, and recall represents the recall rate, but precision and recall are usually in a trade-off relationship, and comprehensive consideration is required for result evaluation. For the RFI recognition task, it is necessary to ensure accuracy, and it is equally important to improve the comprehensiveness of recognition under the condition of ensuring recognition accuracy (usually more than 80%). Therefore, the present invention selects the comprehensive index F2 score as the final evaluation basis. The F2 score is a weighted harmonic mean of precision and recall, which is used to measure the performance of the binary classification model, and is particularly suitable for situations where recall needs to be emphasized. It can be seen from the data that the test results of R-Net perform poorly, and the average result of the comprehensive index F2 is 59.39%; secondly, the F2 index of DSC is in the third echelon, which is 71.86%; finally, the F2 index of U-Net and RFI-Net is not much different, and it shows excellence compared with R-Net. It is worth noting that the MS-TransUNet proposed in the present invention is significantly superior to previous methods in the field of radio astronomy.

[0065] According to the above implementation process, a test picture is randomly selected such as Figure 5 As shown, Figure 6 The corresponding true mask label image is as follows Figure 6As shown in Figure 2, the visualization results of the predicted images obtained by different models are shown in Figure 7-Figure 11 Shown: From Figure 6-Figure 11 It can be seen that although various existing methods can identify RFI, and most RFIs can be identified, the MS-TransUNet network proposed in the present invention shows a better visualization effect than previous methods. Compared with various existing methods, it can predict the shape and position of RFI most comprehensively and accurately.

[0066] In order to explore the rationality of the MS-TransUNet network, the present invention provides the following ablation comparison experiment on the network. As shown in Table 2:

[0067] Table 2 Ablation experiment

[0068] Precision Recall F2 score transUNet 0.8266 0.7779 0.7872 transUNet+1 0.8388 0.7769 0.7885 MS-TransUNet 0.8129 0.8016 0.8039

[0069] As can be seen from Table 2, the original transUNet network performs well in RFI recognition when migrated to the astronomical field, and adding a multi-scale residual position-aware attention module (transUNet+1) to the original network makes better results reflected in the test indicators. On this basis, the branch branch is further introduced, that is, the MS-TransUNet proposed in the present invention is obtained. Although the MS-TransUNet is only a branch branch introduced by a simple convolution kernel combination based on transUNet+1, it brings a good effect to the performance of the network.

[0070] In view of the fact that the characteristics of RFI in astronomical time-frequency images are not unique, some have broadband characteristics and can produce continuous interference in the radio spectrum, and some have burst characteristics, which appear very suddenly in time and have an uncertain duration. In addition, the intensity of RFI is uneven and its distribution in the spectrum is also uneven. The present invention proposes a network structure for identifying RFI based on the improvement of transUNet, called MS-TransUNet. The network combines the ideas of transformer and u-net architecture, uses branch branches as the high-dimensional feature extraction part for extracting interference in time-frequency images, and the multi-scale residual position perception attention module extracts low-dimensional feature information in feature images. The trained model can accurately and comprehensively identify the radio frequency interference in astronomical images in batches, automatically and efficiently. Compared with existing similar methods, it has a faster network convergence speed, reasonable network performance, and can comprehensively and accurately identify RFI. At the same time, the present invention considers the lightweight of the network, sets fewer parameters and a reasonable number of training times. In the case of a limited data set, the present invention expands the data set by rotating and cropping the data online during data preprocessing to ensure the efficiency of learning. This also means that the network of the present invention has the characteristics of being lightweight and highly efficient.

[0071] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A method for identifying radio frequency interference based on astronomical time-frequency images. It is characterized in that include: Acquire astronomical time-frequency images; Construct an MS-TransUNet network model based on multi-scale feature fusion, which consists of an encoder, a branch, and a decoder; use the astronomical time-frequency image as the input of the encoder and the branch to obtain the encoder feature map and the high-dimensional feature map respectively; use the encoder feature map as the input of the decoder, introduce a multi-scale residual position-aware attention module in the decoder to extract feature map information, and concatenate the output of the fourth upsampling in the decoder with the output of the branch to obtain a concatenated feature map, so as to integrate spatial information with high-order semantic information; The spliced ​​feature map is subjected to one convolution operation and then predicted by a segmentation head to be a category of each pixel, so as to output a radio frequency interference mask image; The categories include: interference, non-interference; The Branch specifically includes: performing an addition operation on the outputs of the input astronomical time-frequency image after being respectively operated by a 3×3 convolution kernel and a 1×1 convolution kernel, and then performing batch normalization and ReLU operations to obtain a high-dimensional feature map; The encoder and decoder of the MS-TransUNet network model based on multi-scale feature fusion are based on the encoder and decoder of the TransUNet network. In the encoder, the Transformer block adopts a Transformer layer; in the decoder, a multi-scale residual position-aware attention module is introduced; The encoder comprises a convolution block and a Transformer block; wherein the convolution block extracts local features of the input astronomical time-frequency image to obtain local feature maps of radio frequency interference targets of three different scales; the Transformer block encodes the local feature map of the radio frequency interference target of the third scale and extracts features from it to obtain an encoder feature map; The decoder introduces a multi-scale residual position-aware attention module after the first upsampling; The multi-scale residual position-aware attention module specifically comprises: extracting feature information of the input image of the multi-scale residual position-aware attention module at different scales by setting "expanded convolution with different void rates + average pooling"; splicing the results of the expanded convolution and the average pooling, and then reducing the dimension through a 1×1 convolution; adding the original input image of the multi-scale residual position-aware attention module to the feature map after dimensionality reduction, and then performing a channel shuffling operation on the feature map after batch normalization and Relu operation; performing global average pooling on the feature map after the channel shuffling operation in the horizontal and vertical directions respectively to obtain two feature maps, splicing the feature maps in the horizontal and vertical directions, and then performing 1×1 convolution, batch normalization and nonlinear activation processing, and dividing the processed feature map into two parts in the channel dimension: respectively passing through 1×1 convolution conv, and then passing through Sigmoid The activation function obtains the weight; the "output obtained by multiplying the feature map after the channel shuffle operation by the two weights" and the "output of adding the original input image of the multi-scale residual position-aware attention module multiplied by the preset coefficient and the feature map after the channel shuffle operation multiplied by the preset coefficient" are added as the output of the multi-scale residual position-aware attention module.

2. The method for identifying radio frequency interference based on astronomical time-frequency images according to claim 1, It is characterized in that The MS-TransUNet network model based on multi-scale feature fusion also includes training and testing steps, specifically: The acquired astronomical time-frequency images are annotated and corrected to produce real mask label images, thereby constructing a dataset; According to the preset ratio, the data set is randomly divided into training set, validation set and test set; The training set is expanded by using the Transformer enhancement technology to obtain an expanded training set; The MS-TransUNet network model based on multi-scale feature fusion is trained and verified according to the expanded training set and verification set to obtain the trained MS-TransUNet network model based on multi-scale feature fusion; The astronomical time-frequency images in the test set are used as the input of the trained MS-TransUNet network model based on multi-scale feature fusion to obtain the radio frequency interference mask image.

3. A radio frequency interference identification system based on astronomical time-frequency images, It is characterized in that include: An acquisition module is used to acquire astronomical time-frequency images; The recognition module is used to construct a MS-TransUNet network model based on multi-scale feature fusion, which is composed of an encoder, a branch, and a decoder; the astronomical time-frequency image is used as the input of the encoder and the branch to obtain the encoder feature map and the high-dimensional feature map respectively; the encoder feature map is used as the input of the decoder, and a multi-scale residual position-aware attention module is introduced in the decoder to extract feature map information, and the output of the fourth upsampling in the decoder is spliced ​​with the output of the branch to obtain a spliced ​​feature map, so as to integrate spatial information with high-order semantic information; The spliced ​​feature map is subjected to one convolution operation and then predicted by a segmentation head to be a category of each pixel, so as to output a radio frequency interference mask image; The categories include: interference, non-interference; The Branch specifically includes: performing an addition operation on the outputs of the input astronomical time-frequency image after being respectively operated by a 3×3 convolution kernel and a 1×1 convolution kernel, and then performing batch normalization and ReLU operations to obtain a high-dimensional feature map; The encoder and decoder of the MS-TransUNet network model based on multi-scale feature fusion are based on the encoder and decoder of the TransUNet network. In the encoder, the Transformer block adopts a Transformer layer; in the decoder, a multi-scale residual position-aware attention module is introduced; The encoder comprises a convolution block and a Transformer block; wherein the convolution block extracts local features of the input astronomical time-frequency image to obtain local feature maps of radio frequency interference targets of three different scales; the Transformer block encodes the local feature map of the radio frequency interference target of the third scale and extracts features from it to obtain an encoder feature map; The decoder introduces a multi-scale residual position-aware attention module after the first upsampling; The multi-scale residual position-aware attention module specifically comprises: extracting feature information of the input image of the multi-scale residual position-aware attention module at different scales by setting "expanded convolution with different void rates + average pooling"; splicing the results of the expanded convolution and the average pooling, and then reducing the dimension through a 1×1 convolution; adding the original input image of the multi-scale residual position-aware attention module to the feature map after dimensionality reduction, and then performing a channel shuffling operation on the feature map after batch normalization and Relu operation; performing global average pooling on the feature map after the channel shuffling operation in the horizontal and vertical directions respectively to obtain two feature maps, splicing the feature maps in the horizontal and vertical directions, and then performing 1×1 convolution, batch normalization and nonlinear activation processing, and dividing the processed feature map into two parts in the channel dimension: respectively passing through 1×1 convolution conv, and then passing through Sigmoid The activation function obtains the weight; the "output obtained by multiplying the feature map after the channel shuffle operation by the two weights" and the "output of adding the original input image of the multi-scale residual position-aware attention module multiplied by the preset coefficient and the feature map after the channel shuffle operation multiplied by the preset coefficient" are added as the output of the multi-scale residual position-aware attention module.

4. A processor, It is characterized in that The processor is used to execute an operation, wherein the operation includes executing the radio frequency interference identification method based on astronomical time-frequency images as described in any one of claims 1 to 2.

5. A computer-readable storage medium, It is characterized in that Used to store computer programs; wherein, when the computer program is executed by a processor, the steps of the radio frequency interference identification method based on astronomical time-frequency images as described in any one of claims 1 to 2 are implemented.